Formalizing Neurath's ship: Approximate algorithms for online causal learning.

Area of Science:

• Cognitive Science
• Machine Learning
• Philosophy of Science

Background:

• Higher-level cognition relies on learning world models.
• Exact Bayesian inference for causal models is computationally expensive.
• Cognitive causal learning processes are likely approximate.

Purpose of the Study:

• To propose an algorithmic-level model for approximate causal structure learning.
• To model how learners choose informative interventions under computational constraints.
• To provide a computational framework inspired by Neurath's ship and machine learning.

Main Methods:

• Developed a model representing a single global hypothesis updated locally with evidence.
• Proposed a scheme for selecting informative interventions.
• Analyzed data from 3 experiments to validate the model.

Main Results:

• The proposed model offers a computationally tractable approach to causal structure learning.
• The model provides insights into how learners select interventions to improve causal understanding.
• Experimental results support the proposed algorithmic-level model.

Conclusions:

• Approximate causal learning is essential for higher-level cognition.
• The proposed model captures key aspects of human causal structure learning and intervention selection.
• This work bridges computational approaches in machine learning and cognitive science for understanding causal inference.